PhD in complex environmental systems

Joining:https://wwwfr.uni.lu/formations/fstm/doctoral_programme_in_computational...

PhD candidate on Machine Learning applied to Critical Zone sciences

Terrestrial environment can now be studied in the frame of the Critical Zone (CZ) concept, which naturally generates interdisciplinarity (geology, soil science, biology, ecology, geochemistry, hydrology, geomorphology, atmospheric science, and many more). CZ encompasses the thin outer layer of Earth’s surface extending from the top of the vegetation canopy down to the subsurface depths of fresh groundwater. And interdisciplinarity is more than relevant to study the complex processes involved in the transformation of rock and biomass into soil that in turn sustains most aboveground terrestrial life including Human beings.

Quantitative approaches are needed to predict how CZ evolve naturally and under human impact. Assessment of global change effects on environmental management requires dealing with multi-dimensional information. Such approach generates on the one hand new knowledge, gained via development of interdisciplinary approaches, and on the other hand higher level of uncertainties on the environmental change predictions. One of the possibility to solve this paradox is to develop new data driven processes to model environmental changes. When CZ is observed over short timeframes, the response to perturbation is often linear, but over long timeframes, nonlinear responses emerge, because the system is driven by a complexity of coupled processes. As a consequence, the models that are used to simulate CZ evolution are extraordinarily difficult to develop because they must include earth materials, fluids, biota and human behaviour over extended lengths of time. Moreover, the non-stationarity of environmental evolution driven by climate change effect makes model predictions based on historical data series more uncertain.

Machine Learning represents a pragmatic breakthrough in making predictions by finding complex structures and patterns in large volumes of data. The development of increasingly sophisticated machine learning techniques, combined with rapid increases in computational ability, has prompted extensive research into advanced methods for data driven environmental processes prediction within CZ in the past decade. Examples can be found in hydrology (watershed runoff, regulation in river flows, groundwater remediation), biogeochemistry (modelling nitrate distributions in aquifers, CO2 sequestration during bedrock weathering), geohydrology (spatially distributed 3D-characterization of regolith, spatial organization of soil physical properties). Artificial neural networks, regression trees, and support vector machines have been shown to be powerful tools for predictive modelling and exploratory data analysis in systems that exhibit complex and non-linear behaviour. However, the Machine Learning models are still heavily criticized for the lack of interpretability of induced models and are often referred to as a black box. As a result of the lack of interpretability, the contribution from such data-driven models to scientific advancement still remains negligible. Incorporating available scientific knowledge and large datasets to guide learning algorithms to generate more reliable and consistent model predictions will be an effective way to improve the explicability of machine learning models used for complex environmental systems.

One of the challenges of using Machine Learning to CZ investigations is to determine the optimal data to acquire. These data are often used to either provide structural information or constrain model parameterization. Measurement optimization is an attempt to balance field and laboratory data quality. One possibility is to develop an efficient and robust approach to optimize the data, with the hope that a similar approach could be extended into various fields of application or to complementary research questions. One of the possibility is to develop a Machine Learning approach based on robust statistics to help frame a predictive modeling.

problem and better understand the data. That statistical contribution can be used to clean and prepare data ready for modelling and can be defined to test and estimate model selection. A challenge in such analyses is to specify the criteria for selecting the proper datasets to combine and construct a predictive model.

contribute to the knowledge advancement on the functioning of CZ-based research by improving the use of large and high-frequency dataseries on specific environmental observatory.

To strengthen the knowledge on CZ evolution and processes modelling in the context of global changes, the University of Luxembourg in collaboration with the Luxembourg Institute of Science and Technology is searching for a high-level candidate to defend PhD project funding on Machine Learning applied to CZ sciences.

You will develop and evaluate Machine Learning models enabling the classification of short-term to long-term datasets recorded in one of the most equipped CZ Observatory for ecohydrological purposes in Luxembourg. This experimental site has been recording environmental data for more than 15 years in different CZ compartments (vegetation, soil, groundwater, streamwater) The research will be carried out in the framework of a 4 years PhD project supported by the Luxembourg National Research Fund (FNR) through its AFR program.

More specifically you would contribute to the project by:

· Developing and coding innovative scientific Deep Learning/Machine Learning algorithms to classify environmental data series (hydro-climatological records, hydrochemistry, tree physiology)

· Acquiring interdisciplinary skills in between computational and environmental sciences

· Contributing to software development, integration, testing and deployment.

· Disseminating and publishing the results in top ranked scientific journals

· Contributing to the development of partnerships and networks at national and international levels.

· Participation in the implementation of technological solutions (proof-of-concepts, prototypes).